Competitive intelligence for clinical trials.¶
Ogur watches clinical trial registries, FDA filings, patent offices, company IR feeds, and the scientific literature — detects competitive changes — and synthesizes them into structured briefings for pharma strategy teams. This site is the technical documentation; for the source code, browse the repository.
First contact.¶
-
Installation
Prereqs, environment variables, seeding the SQLite DB, frontend bootstrap.
-
Architecture
System diagram, data model, engine internals, KIQ + verification gate, dedup invariants.
-
Contributing
Make-target catalogue, code style, recipes for adding sources / agents / analyzers.
What lives where.¶
-
API reference
Every FastAPI endpoint with request/response schema, query parameters, status codes, curl examples.
-
Data sources
Per-source authentication, rate limits, signal types produced, upstream documentation links.
-
Frontend
Vite + React stack, view structure, the Inspector pattern, structured-entity-ref rendering.
-
Testing
Fixture catalogue, the patch-where-used rule, factory functions, the StaticPool pattern.
How we measure.¶
-
Evaluation
Briefing-quality axes, the entity / outcomes eval harnesses, KIQ-shape verification gate, cost-per-run budget.
-
Research
Provenance-first design, agent architecture, BIOPSY benchmark, retrieval seam, competitive positioning.
The visual contract.¶
-
UX specification
Canonical visual + interaction contract. The file that wins when implementation disagrees.
-
Agentic architecture notes
Coordinator/specialist pattern inspired by Claude Code, applied to the synthesis pipeline.
Where Ogur is right now.¶
MVP-1 released 2026-08-13 as v1.1.0 — named client users read auditable,
frozen-pack competitive-landscape reports behind authentication in production,
and leave attributable feedback. The active milestone is MVP-2: the Explore
backend (discovery and landscape generation), currently being scoped. The
contract is docs/product/mvp-scope.md.
| Layer | Status | Scope |
|---|---|---|
| Ingestion | Complete | 16 sources, content-hash dedup |
| Intelligence engine | Complete | Detect → classify (rules + Haiku) → enrich → synthesize → verify; QueryEngine for Q&A |
| Evidence layer | Complete | Entity NER (GLiNER + Claude), outcomes extraction (Haiku tool-use), CT.gov v2 parser, two eval harnesses |
| API | Complete | Auth/session, signals, briefings, comparative evidence, metered ask, explore report/figures, feedback |
| Frontend | Complete (MVP-1) | Vite/React; client build serves the report-pack surfaces, demo surfaces build-gated |
| Delivery | Complete (MVP-1) | Immutable container (packs + content DB), Fly.io, production smoke gate |
| Explore discovery backend | MVP-2 — in definition | Candidate retrieval, discovery jobs, landscape generation |